A month ago I wrote that I have more game ideas than evenings. What I did about it is where this gets embarrassing: I decided that if the evenings were fixed, I'd simply put more into each one. You can probably see the shape of what happened next.
It's about half nine on a Tuesday. The day job finished hours ago, dinner's done, and I've got eight to ten agent sessions open across the studio's projects. I tab into one — and for four seconds I have no idea what it is. Not what it's doing. What it is. Which project, which branch, whether that's a plan I approved or work already committed while I was confused somewhere else.
Four seconds isn't a crisis. But it happens thirty-odd times between dinner and midnight, and by eleven I'm not tired from building anything — I'm tired from re-entering. I've started calling it drAIn, because the work isn't what drains, it's the reloading. Also because I enjoy a bad pun more than is strictly professional.
Everyone talks about the other drain, the ambient one: models change, harnesses change, last month's best practice is this month's footgun. The one that got me is sharper. Every session I'm not looking at is a context I'm still half-holding, and every tab costs an unload and a load — from a low starting number, because this is the second shift and the day job already spent it.
I did not discover this
I got four hundred words into this post believing I'd noticed something. I had not.
In March, Harvard Business Review ran a piece on exactly this, with a better name than mine: brain fry. The underlying study — BCG Henderson Institute, 1,488 US workers — separates the people who use AI from the people who supervise it, and the cost lands on the second group: 14% more mental effort, 12% more mental fatigue, 19% more information overload. It also separates brain fry from burnout deliberately — burnout is chronic, brain fry is acute. It's what Tuesday night feels like, and it's gone by Wednesday morning, at which point I go to a job that has nothing to do with any of this and come home restocked, ready to do it again. Recovery, it turns out, is not the same as learning.
Then there's Addy Osmani, who wrote the sentence I wish I'd had six months ago: "Your cognitive bandwidth doesn't parallelize." He names the ambient anxiety tax — background vigilance about what's quietly going wrong in the threads you aren't watching — and comprehension debt, understanding falling behind what the agents generate. His workable ceiling is three to four threads. I was running ten, at night, after a full day of work, and had concluded that what I needed was a better keyboard shortcut.
The mechanism underneath is older than any of it: Sophie Leroy's attention residue, published in 2009, found that part of your attention stays stuck on the task you left, especially when it was unfinished. Which describes ten agent sessions exactly. They are all unfinished. That's what makes them sessions.
There's also a much-shared ladder of how many sessions each tier of tooling supports, which flattered me enough that I checked it: a reasoned estimate on a personal blog, apparently written by a model rather than a person. Being told by a language model that I'm operating at tier three is precisely the compliment I should distrust.
The one useful thing: markdown is a serial channel
For months I tried to fix the workload — fewer sessions, tighter briefs, better scoping. That helps, and I do it now, but it treats the symptom, because the thing being overloaded isn't my capacity to work. It's my capacity to orient. The agents parallelised generation; my input channel stayed serial, and reading does not get wider at half past ten.
Which turned my attention to the format of what I was reading rather than the amount. Every agent hands me markdown, and as a status report markdown has one fatal property: you have to read it, start to end, at the speed of language. A ten-repo report is ten paragraphs, and you find out the fourth one is on fire when you reach the fourth paragraph.
That isn't how seeing works. Colour, size and position are perceived pre-attentively — they
land before you've decided to look. You know which bar is tallest before you read the axis. So
when you are the bottleneck, the legibility of the status report is itself the
intervention. Not a nicety on top of the fix. The fix. A repo in trouble should look like a
repo in trouble before you've read a word about it — which is a grand way of describing a
<div> as wide as the mess is big:
<!-- the whole trick: the bar's width is the size of the unfinished work -->
<div class="meter"><div class="fill" style="width:99%"></div></div>
So I built a dashboard instead of closing some tabs
Let the record show that a reasonable person, on discovering they were running twice the sustainable number of AI sessions, might have run fewer AI sessions. I built an eleventh tool to supervise the other ten. I contain multitudes, most of them unfinished.
It's a Claude Code skill called sitrep. I ask what's pending; it scans every git repo in the
workspace, works out what each piece of unfinished work actually is, and renders the lot as
one self-contained HTML page. The dirty-tree meter scales with the change, so ninety
half-finished files look like ninety half-finished files. Stashes get a dashed box, because
dashed reads as detached — orphaned work, attached to nothing, the easiest thing in git to
lose forever. And the stat blocks up top are the five-second version, which most nights is all
I read.
It's read-only, deliberately: it tells me what needs a decision, and I make the call. My judgement is the scarce resource, so handing it to the tool would solve the wrong half — and besides, I've seen what I write at eleven at night. I don't want a robot acting on it unsupervised either.
The difference isn't that I ship more. It's that the first twenty minutes of every evening stopped being an archaeology dig, and I stopped losing things in stashes.
Take it
Download the sitrep skill (10 KB)
Unzip it into ~/.claude/skills/, make the scan script executable, and ask for a sitrep. It
defaults to ~/Development and takes a path argument if yours lives elsewhere. MIT licensed.
I've contradicted myself here and would rather point at it than hope nobody notices: in Meet the team I wrote that you can't download the studio's tools, and that stands — Forge, Herald, Odin and the rest stay internal. This isn't one of them. It's a markdown file and a shell script: not a tool that builds the games, but the thing that lets me survive running the tools that do.
What it doesn't fix
It doesn't raise my ceiling. Comprehension debt is still sitting right there — I still can't genuinely understand the output of ten sessions, and pretending otherwise is how you ship code nobody read. A sitrep lowers the cost of orientation, not the cost of understanding, and those are different currencies. What it bought me isn't throughput; it's fewer things quietly rotting on a branch I'd forgotten existed. Less a productivity gain than a reduction in the number of ways I can disappoint myself, which at this stage I'll take.
So if your own number is creeping past four, don't ask how to run more sessions. Ask what your status report is asking of you, and whether it has any business asking it. Then, unlike me, consider just closing a couple of tabs.
More soon — RSS if you'd like to watch me solve problems I created.